Idea: vector_native contract encoding (#70)

[ariaxhan]

vector_native: Compressed Agent-to-Agent Communication

Issue: #70 | Priority: P2 | Origin: Crystal OS vector_native protocol

The Problem

KERNEL agents pass contracts, checkpoints, and handoffs in prose. A tier-3 cycle (orchestrator → surgeon → adversary → reviewer) can burn ~1500 tokens on coordination overhead alone. That's context window space that should carry code understanding, not prose.

The Proposal

Encode AgentDB communications in vector_native — a symbolic protocol that achieves 83-85% compression:

Before (50 tokens):

Contract: Fix the authentication timeout in middleware/auth.ts.
Files in scope: middleware/auth.ts, middleware/session.ts.
Approach: Increase timeout from 5s to 30s, add retry logic.

After (12 tokens):

●CONTRACT|goal:fix_auth_timeout|files:[middleware/auth.ts,middleware/session.ts]|approach:increase_timeout+add_retry|constraints:[preserve_session,tests_pass]

Token Savings

Communication	Current	vector_native	Savings
Simple contract	50 tokens	12 tokens	76%
Complex contract	300 tokens	45 tokens	85%
Full tier-3 cycle	~1500 tokens	~250 tokens	83%

Those 1250 saved tokens = more code in the agent's context window = better decisions.

Key Design Choice

Hybrid encoding: vector_native in AgentDB records (machine-to-machine), human-readable prose in _meta/ artifacts (human-readable). Both generated from the same structured data.

For Agents Working On This

Read the full spec in #70. The vector_native repo already exists at CodingVault/vector-native/. This is integration work, not invention. Start with contracts only, expand to checkpoints and handoffs.